About AI Speed Index
Fast AI models, measured.
AI Speed Index is an independent testing publication. We measure fast AI models — latency, throughput, cost, reliability, privacy, and how they actually behave inside real workflows — and we publish the numbers, the method, and the places where things went wrong.
We start with GLM 5.3 Fast. We are not exclusive to it. GPT, Claude, Gemini, Llama, and whatever ships next month are all fair game for comparison, because a speed claim only means something when it sits next to a competitor.
Why this site exists
If you are a product manager, a marketer, an independent developer, an ops lead, or a founder on a small team, you have probably done this at least once: opened six browser tabs, skimmed four pricing pages, watched two YouTube reviews, and still had no idea which model would finish your actual work faster without quietly costing you three times as much.
The information exists. It is just scattered across provider documentation, benchmarks that measure things you will never do, and demos that nobody can reproduce.
AI Speed Index is the site we wanted to read: transparent methods, repeatable prompts, U.S.-relevant pricing, real workflow tests, and public limitations.
Who is behind it
My name is Jordan Lee. I spent nine years as a SaaS product manager before becoming an independent AI product tester and consultant. I live in Austin, Texas.
I am not employed by a model company. I am not an AI research scientist, and I do not pretend to be one. The perspective here is the one I actually have: a product operator who buys, deploys, and troubleshoots this stuff. What does the model do? How consistently does it do it? What does it cost? When should a real person use it — and when should they walk away?
Tests are run from a home office on a MacBook Pro over a 1G fiber connection, under U.S. Central Time conditions. That is not a data center. It is also what a lot of readers are actually working with, which is the point.
How we test
Every substantial test documents the same things:
Model name and exact version
Date and time of testing
Provider and API endpoint
Geographic and regional test conditions
Prompt set or task description
Input and output token counts
Number of runs
First-token latency
Total completion time, with average, median, and range
Tokens per second
Failure rate and retry behavior
Pricing assumptions
Known limitations
Whether the provider sponsored, supplied, or influenced the test
And every claim on this site is labeled as one of four things:
Measured — we observed it directly
Provider claim — the vendor said it; we have not confirmed it
Editorial judgment — my interpretation, and you can disagree
Open question — not yet tested, not yet verified
If you want the full detail, read How We Test AI Model Speed. It is long, and it is supposed to be.
What we will not do
Publish investment-style hype about AI companies.
Claim that one benchmark proves universal superiority.
Present an unrepeatable demo as scientific evidence.
Pretend to have insider access or credentials that do not exist.
Hide a failed test because it weakens the headline.
Some tests will come back inconclusive. Some will produce a result that makes the previous article look a little silly. Both get published, with a note explaining what changed.
On independence
Nobody pays for a verdict here. If a model provider ever sponsors, supplies, or otherwise touches a test, that fact appears at the top of the test — not in a footnote, and not in a word like "partnered."
We publish corrections when we get something wrong. If you find an error in the data, the method, or the math, tell us and we will fix it and say what we changed.
What we are building toward
Individual articles bring readers in through search. The long-term asset is the Fast Model Index — a recurring ranking that scores models on speed, cost, output quality, reliability, and workflow fit, with the raw data published alongside it so you can reweight it yourself.
Our current weighting: Speed 25%, Cost 20%, Output quality 25%, Reliability 20%, Workflow fit 10%. You are allowed to think that is the wrong weighting. That is exactly why the underlying numbers are public.
One boundary
I write about my work. I do not write about my family. You will not find my neighborhood, my kid's name or school, or my spouse's employer on this site, and that is deliberate.
The short version
Fast is a feature. Reliable is the product.
If a model is quick and falls over under load, it is a demo. If it is quick, boring, and still standing after ten thousand requests, it is a tool. This site is about telling those two apart — with a stopwatch, a spreadsheet, and no patience for hype.
Start here: The Fast Model Index · Latest test · Testing method